What it is
Qwen3.5-397B-A17B was the first open-weight model in Alibaba's Qwen3.5 series, released on February 16, 2026. Alibaba has since released Qwen3.6 open models in April and Qwen3.8 in August 2026. It is a mixture-of-experts model with 397 billion parameters, of which 17 billion are used for each token. It mixes a fast linear-attention design (Gated DeltaNet) with standard attention, which Alibaba says gives 8.6 times the decoding speed of Qwen3-Max at 32K context and 19 times at 256K.
It reads text, images and video and writes text. The native context is 262,144 tokens, extendable to about 1M with YaRN, a standard context-stretching method. It covers 201 languages and dialects and thinks before answering by default. The hosted version, Qwen3.5-Plus on Alibaba Cloud Model Studio, adds a 1M-token default context and built-in tools.
The weights are under Apache 2.0. Smaller Qwen3.5 models from 0.8B to 122B followed in late February and early March 2026.
Our review
Qwen3.5-397B-A17B was a strong open model in February and is an expensive one in October. Artificial Analysis scores it 18 on its Intelligence Index, just under the median of 19 for open models above 150 billion parameters, and calls it 'particularly expensive' for its class at $0.60 input and $3.60 output per million tokens. Alibaba's own card, which is self-reported, still reads well: 76.4 on SWE-bench Verified, 88.4 on GPQA, 85.0 on the MMMU image test and 62.2 on OSWorld-Verified.
Alibaba has replaced it. Qwen3.6-35B-A3B and Qwen3.6-27B arrived in April, and in August Qwen3.8 brought a 2.4-trillion-parameter flagship and a 27B model. The flagship ships under a custom Qwen3.8-Max licence, while the 27B is Apache 2.0, so the 397B remains the largest Qwen with plain Apache terms. That is its narrowest and most durable argument.
The hosting maths cuts against it. Qwen3.5-Plus, the hosted version, is listed on OpenRouter at $0.30 input and $1.80 output with a 1M-token context, while the cheapest of ten OpenRouter hosts for the open weights, Alibaba's own endpoint, charges $0.39 and $2.34 with 262K. Self-hosting only pays when the weights must sit in your own building. The FP8 checkpoint is about 406 GB and Alibaba's serving examples assume eight GPUs; Unsloth's 4-bit quant, at 214 GB, fits a 256 GB Mac or a single 24 GB GPU with 256 GB of system RAM at about 25 tokens a second.
Two behaviours catch teams out. Thinking is on by default, the /think and /nothink switches are not supported, and the card asks for at least 128K of context to keep reasoning intact, which constrains memory on smaller servers. Past 262K, open-source frameworks use static YaRN, which the card warns can hurt short-prompt quality, so long-context work belongs on the hosted Plus. In its favour, Artificial Analysis counted 95M output tokens across its index against a 140M median, which makes it terse for a reasoning model even at 55.8 tokens a second on Alibaba's API.
Where it earns its place
An ML platform team at a European bank with an eight-GPU H100 node
Run a multimodal model on-premises under a licence legal has already approved
Serve the 406 GB FP8 checkpoint with vLLM at tensor-parallel 8, the qwen3 reasoning parser and the qwen3_coder tool parser. Apache 2.0 has no thresholds to track; the cost is the node, and the card advises keeping at least 128K of context.
A solo developer with a 256 GB Mac Studio
Run a large open model locally for private coding and document work
Unsloth's UD-Q4_K_XL quant is about 214 GB and loads on a 256 GB M3 Ultra; Unsloth cites a third-party test where it lost under a point against the full weights. Expect single-user speeds, not a team server.
A support-operations manager at a SaaS company answering 1M tickets a month
Draft multilingual replies without paying flagship prices
At 1,500 input and 500 output tokens a reply, Alibaba's endpoint on OpenRouter ($0.39/$2.34) comes to about $1,760 a month, and hosted Qwen3.5-Plus ($0.30/$1.80) to about $1,350. Reasoning tokens bill as output, so turn thinking off for routine replies.
Before you commit
- Setup
- Download from Hugging Face or ModelScope (about 807 GB in BF16, 406 GB for the FP8 checkpoint) and serve with main-branch vLLM or SGLang; the card's examples use eight GPUs. Or call qwen/qwen3.5-397b-a17b on OpenRouter.
- Integrations
- OpenAI-style chat APIs on Alibaba Cloud Model Studio and ten OpenRouter hosts; tool calling through the qwen3_coder parser; KTransformers, llama.cpp, MLX and Transformers for local runs.
- Data and privacy
- Apache 2.0 allows commercial use, fine-tuning and redistribution with no revenue or user limits. Self-hosting keeps prompts in-house; on OpenRouter it is not available through EU in-region routing, and each host sets its own retention.
- Cost at scale
- OpenRouter hosts charge $0.39 to $0.75 input and $2.34 to $4.50 output per 1M tokens; hosted Qwen3.5-Plus is $0.30/$1.80. Self-hosting needs about 406 GB for FP8 (an eight-GPU 80 GB server), 214 GB at 4-bit or 180 GB at 3-bit.
- Limits
- 262,144-token native context, about 1M with YaRN factor 4; output up to 32,768 tokens by default and 81,920 for hard maths and code. Some hosts cut context to 128K or 131K, and Alibaba's API runs at 55.8 tokens a second per Artificial Analysis.
What's good
- Apache 2.0 with no revenue or user caps
- Text, image and video input across 201 languages
- Terse for a reasoning model: 95M output tokens on Artificial Analysis's index vs a 140M median
- 4-bit quant fits in 214 GB of memory
What's not
- Artificial Analysis Intelligence Index of 18, below the median for its size
- Hosted Qwen3.5-Plus undercuts every open-weight host on OpenRouter
- Superseded by Qwen3.6 and Qwen3.8
- Beyond 262K relies on static YaRN, which can hurt short-text quality
Best for
Teams that need the largest Apache 2.0 Qwen on their own hardware and already have a multi-GPU server or a 256 GB machine to run it.
Alternatives
Qwen3.5-397B-A17B: common questions
- Is Qwen3.5-397B-A17B worth it?
- Still a capable Apache 2.0 workhorse, but Alibaba now sells the same model hosted for less than any third party charges to run the open weights. We rate it 3.6 out of 5.
- How much does Qwen3.5-397B-A17B cost?
- Free weights under Apache 2.0; Qwen3.5-Plus on Alibaba Cloud and third-party hosting billed separately.
- What are the downsides of Qwen3.5-397B-A17B?
- Artificial Analysis Intelligence Index of 18, below the median for its size; hosted Qwen3.5-Plus undercuts every open-weight host on OpenRouter; superseded by Qwen3.6 and Qwen3.8; beyond 262K relies on static YaRN, which can hurt short-text quality.
- Who is Qwen3.5-397B-A17B best for?
- Teams that need the largest Apache 2.0 Qwen on their own hardware and already have a multi-GPU server or a 256 GB machine to run it.
- How much does Qwen3.5-397B-A17B cost for a team or in production?
- OpenRouter hosts charge $0.39 to $0.75 input and $2.34 to $4.50 output per 1M tokens; hosted Qwen3.5-Plus is $0.30/$1.80. Self-hosting needs about 406 GB for FP8 (an eight-GPU 80 GB server), 214 GB at 4-bit or 180 GB at 3-bit.
- Is Qwen3.5-397B-A17B safe for company data?
- Apache 2.0 allows commercial use, fine-tuning and redistribution with no revenue or user limits. Self-hosting keeps prompts in-house; on OpenRouter it is not available through EU in-region routing, and each host sets its own retention.
- What are the best Qwen3.5-397B-A17B alternatives?
- The closest alternatives we have reviewed are GLM-5.3-Flash, Qwen3.8-Max, DeepSeek-V4.1-Flash and Mistral Small 4.
Answers checked Oct 22, 2026.